Reduce AI context inside WSO2. Keep the evidence path.

Preliminary observed result

A reason to evaluate—not a universal performance claim.

In a 60-call comparison over 20 frozen, synthetic 20K–100K-token workloads, run once on 2026-08-17, the Maha path reduced provider input tokens by 98.84% and observed model cost by 98.20% relative to baseline. All 60 calls completed without retries or failures.

*The Prompt Compressor result is specific to WSO2 AI Gateway 1.1.0, Prompt Compressor 0.9.0, and a 0.55 retained ratio. It should not be generalized before WSO2 or a customer confirms that configuration reflects the intended production setup. Provider pricing assumptions and the full frozen configuration are recorded in the reproduction manifest.

Which retention score this is. “Required facts retained” above is the path-blinded semantic rubric applied to the returned answers. A second, stricter scorer measures exact evidence-span containment on the same answers and reports 15 / 60 for the baseline, 0 / 60 for the Prompt Compressor, and 24 / 60 for Maha, because it scores a correct paraphrase as a miss. Both scores are published per workload in the evidence artifact. Latency is one observation per call on a single run, not a percentile over repeated runs.

Reproduce before trusting

The evaluation method is inspectable.

The public package pins the gateway and policy versions, retained ratio, model, corpus digest, failure behavior, and zero-retry rule. Its default one-command run is a dry run and makes no provider calls.

The evidence package

Four documents, each served from this site and each stating its own scope. Read them before the call rather than after it.

• WSO2 evaluation evidence artifact ↗

  Every call in the frozen three-path evaluation as its own row, with the aggregates re-derived from those rows.

• Sample Context-Control Evidence Assessment ↗

  The deliverable a customer receives, produced from a synthetic corpus. Not a customer result.

• Security and data-boundary one-pager ↗

  What the compiler and its WSO2 interceptor do with context data, with each statement traced to committed source.

• MCRB-1 dense retriever baseline ↗

  An embedding-retrieval baseline on the frozen MCRB-1 cohort, published beside the v1 results it is compared with.

The retention benchmark publishes its v1 results alongside the dense baseline, so the comparison can be read in full rather than in the direction that flatters us.

Check the observed result yourself

The table above is not typed into this page. It is read from a committed artifact that carries every one of the 60 calls as its own row, and whose totals are re-derived from those rows each time it is loaded — a hand-edited total fails validation instead of rendering. Verify the artifact you are reading is the one described here:

Evidence artifact SHA-256

78a822dd3c85821a98ca01262b02c12b5c42bc3c1433515909d3a0c5554aa3f1

Frozen corpus label digest

a6b95c81e981fcda20576046354105994bc8846e2b6da8a4e77c754a7085a8c9

Source checkpoint SHA-256 (not published)

sha256:40cb6956b0b732918c1bf6d63852d1076505c0348dd195a86384a5a935f4d0f3

The run’s primary evidence — the durable checkpoint and the path-blinded adjudication — is retained outside this repository because both carry the model’s answer text for every call. They are identified above by digest so a reviewer under NDA can be handed the exact bytes this artifact was derived from and re-derive it with npm run generate:wso2-live-evidence. The sanitized single-workload trace remains published separately: it is one representative call, not evidence for the aggregate.

Start with evidence

Bring one expensive or audit-sensitive context workflow.

Send the approximate input size, request volume, current WSO2 AI Gateway version, and the facts or citations that must survive. Maha will reply with a bounded scope—or say plainly if the workflow is not a fit.